import os import sys import yaml import argparse import pandas as pd import numpy as np import matplotlib.pyplot as plt import seaborn as sns import joblib def load_config(config_path): with open(config_path, "r", encoding="utf-8") as f: return yaml.safe_load(f) def main(): parser = argparse.ArgumentParser(description="Generate visualization plots for AI text detection.") parser.add_argument("--config", default="configs/config.yaml", help="Path to config file") args = parser.parse_args() config = load_config(args.config) output_dir = config["paths"]["output_dir"] reports_dir = config["paths"]["reports_dir"] models_dir = config["paths"]["models_dir"] plots_dir = os.path.join(reports_dir, "plots") os.makedirs(plots_dir, exist_ok=True) os.makedirs(output_dir, exist_ok=True) # Load predictions preds_path = os.path.join(output_dir, "recent_debates_predictions_v2.csv") if not os.path.exists(preds_path): preds_path = os.path.join(output_dir, "recent_debates_predictions.csv") if not os.path.exists(preds_path): print(f"Error: Predictions file not found. Please run inference first.") sys.exit(1) df = pd.read_csv(preds_path) df["date"] = pd.to_datetime(df["date"]) df["year"] = df["date"].dt.year # Set premium aesthetic style sns.set_theme(style="whitegrid") plt.rcParams["figure.facecolor"] = "#fbfbfb" plt.rcParams["axes.facecolor"] = "#ffffff" plt.rcParams["font.sans-serif"] = ["DejaVu Sans", "Helvetica", "Arial"] plt.rcParams["font.family"] = "sans-serif" colors_palette = ["#3f51b5", "#e91e63", "#00bcd4", "#ff9800", "#4caf50", "#9c27b0"] sns.set_palette(colors_palette) # --- Plot 1: AI Suspicion Over Time (2004-2026, Weekly & Log Scale) --- print("Generating Plot 1: AI Suspicion Over Time (Weekly & Log Scale)...") plt.figure(figsize=(12, 6)) # Calculate weekly starting date df["week_start"] = df["date"] - pd.to_timedelta(df["date"].dt.weekday, unit='D') weekly_avg = df.groupby("week_start")["prob_ai"].agg(["mean", "count"]).reset_index() weekly_avg.columns = ["week_start", "prob_ai_mean", "speech_count"] # Epsilon for log scale weekly_avg["prob_ai_plot"] = weekly_avg["prob_ai_mean"] + 1e-4 sns.lineplot(data=weekly_avg, x="week_start", y="prob_ai_plot", color="#3f51b5", linewidth=1.5, alpha=0.8) plt.scatter(weekly_avg["week_start"], weekly_avg["prob_ai_plot"], s=weekly_avg["speech_count"]*2, color="#3f51b5", alpha=0.6, label="Moyenne hebdo (taille = nb de discours)") # Add a red dashed line at 2022-11-30 to show release of ChatGPT chatgpt_release = pd.to_datetime("2022-11-30") plt.axvline(x=chatgpt_release, color="#e91e63", linestyle="--", alpha=0.8, linewidth=1.5, label="Sortie de ChatGPT (Fin 2022)") plt.title("Évolution hebdomadaire du score moyen de suspicion d'IA (Échelle Log, 2004-2026)", fontsize=14, fontweight="bold", pad=15) plt.xlabel("Date", fontsize=12) plt.ylabel("Score de suspicion moyen (Probabilité d'IA + 1e-4)", fontsize=12) # Log scale plt.yscale("log") plt.ylim(0.5e-4, 1.2) plt.yticks([1e-4, 1e-3, 1e-2, 1e-1, 1.0], ["0.0001 (Humain)", "0.001", "0.01", "0.1", "1.0 (IA)"]) plt.legend(loc="upper left", frameon=True) plt.tight_layout() plot1_path = os.path.join(plots_dir, "suspicion_over_time.png") plt.savefig(plot1_path, dpi=300) plt.savefig(os.path.join(output_dir, "suspicion_over_time.png"), dpi=300) plt.close() # --- Plot 2: Probability Distribution Histogram --- print("Generating Plot 2: Probability Distribution...") plt.figure(figsize=(10, 5.5)) if "actual_label" in df.columns: # If we have actual labels, plot separate distributions sns.kdeplot(data=df[df["actual_label"] == 0], x="prob_ai", fill=True, label="Discours Humain", color="#3f51b5", alpha=0.5, bw_adjust=0.5) sns.kdeplot(data=df[df["actual_label"] == 1], x="prob_ai", fill=True, label="Discours IA", color="#e91e63", alpha=0.5, bw_adjust=0.5) plt.title("Distribution des scores de suspicion (Humain vs Synthétique)", fontsize=14, fontweight="bold", pad=15) else: sns.histplot(data=df, x="prob_ai", bins=30, kde=True, color="#3f51b5", alpha=0.7) plt.title("Distribution générale des scores de suspicion d'IA", fontsize=14, fontweight="bold", pad=15) plt.xlabel("Score de suspicion (Probabilité d'IA)", fontsize=12) plt.ylabel("Densité", fontsize=12) plt.xlim(-0.05, 1.05) plt.legend(loc="upper right", frameon=True) plt.tight_layout() plot2_path = os.path.join(plots_dir, "probability_distribution.png") plt.savefig(plot2_path, dpi=300) plt.savefig(os.path.join(output_dir, "probability_distribution.png"), dpi=300) plt.close() # --- Plot 3: Suspicion Score by Political Party --- print("Generating Plot 3: Suspicion Score by Party...") plt.figure(figsize=(10, 5.5)) party_avg = df.groupby("party")["prob_ai"].mean().reset_index().sort_values(by="prob_ai", ascending=False) sns.barplot(data=party_avg, x="prob_ai", y="party", palette="viridis") plt.axvline(x=df["prob_ai"].mean(), color="#e91e63", linestyle=":", alpha=0.8, linewidth=1.5, label="Moyenne générale") plt.title("Score moyen de suspicion d'IA par groupe politique", fontsize=14, fontweight="bold", pad=15) plt.xlabel("Score de suspicion moyen", fontsize=12) plt.ylabel("Groupe politique", fontsize=12) plt.xlim(0, max(party_avg["prob_ai"].max() + 0.05, 0.5)) plt.legend(loc="lower right", frameon=True) plt.tight_layout() plot3_path = os.path.join(plots_dir, "suspicion_by_party.png") plt.savefig(plot3_path, dpi=300) plt.savefig(os.path.join(output_dir, "suspicion_by_party.png"), dpi=300) plt.close() # --- Plot 4: Suspicion Score by Document Type --- print("Generating Plot 4: Suspicion Score by Document Type...") plt.figure(figsize=(10, 5.5)) doc_avg = df.groupby("document_type")["prob_ai"].mean().reset_index().sort_values(by="prob_ai", ascending=False) sns.barplot(data=doc_avg, x="prob_ai", y="document_type", palette="rocket") plt.axvline(x=df["prob_ai"].mean(), color="#3f51b5", linestyle=":", alpha=0.8, linewidth=1.5, label="Moyenne générale") plt.title("Score moyen de suspicion d'IA par type de document", fontsize=14, fontweight="bold", pad=15) plt.xlabel("Score de suspicion moyen", fontsize=12) plt.ylabel("Type de document", fontsize=12) plt.xlim(0, max(doc_avg["prob_ai"].max() + 0.05, 0.5)) plt.legend(loc="lower right", frameon=True) plt.tight_layout() plot4_path = os.path.join(plots_dir, "suspicion_by_doc_type.png") plt.savefig(plot4_path, dpi=300) plt.savefig(os.path.join(output_dir, "suspicion_by_doc_type.png"), dpi=300) plt.close() # --- Plot 5: Top 10 Coefficients for IA vs Human --- # Load model to extract coefficients/importances model_pkg_path = os.path.join(models_dir, "best_detector_v2.pkl") is_v2 = os.path.exists(model_pkg_path) if not is_v2: model_pkg_path = os.path.join(models_dir, "best_detector.pkl") if os.path.exists(model_pkg_path): print(f"Generating Plot 5: Top Features from {os.path.basename(model_pkg_path)}...") pkg = joblib.load(model_pkg_path) if is_v2: xgb_raw = pkg["xgb_raw"] stylometric_cols = pkg["stylometric_cols"] importances = xgb_raw.feature_importances_ COLS_MAP_V2 = { 'num_chars': "Nombre de caractères", 'num_words': "Nombre de mots", 'num_sentences': "Nombre de phrases", 'avg_sentence_len': "Longueur moyenne des phrases", 'std_sentence_len': "Écart-type longueur des phrases", 'slv_normalized': "Complexité lexicale (SLV)", 'avg_word_len': "Longueur moyenne des mots", 'ratio_long_words': "Ratio de mots longs (>6 chars)", 'vocabulary_diversity': "Diversité lexicale (TTR)", 'hapax_ratio': "Ratio d'Hapax (mots uniques)", 'yules_k': "Richesse lexicale (Yule's K)", 'maas_index': "Indice Maas", 'information_entropy': "Entropie de l'information", 'brunet_w': "Indice de Brunet W", 'ratio_punctuation': "Ratio de ponctuation", 'freq_uppercase': "Fréquence des majuscules", 'freq_digits': "Fréquence des chiffres", 'connector_ratio': "Ratio de connecteurs logiques", 'connector_diversity': "Diversité des connecteurs", 'repetition_ratio': "Ratio de répétitions lexicales", 'stopword_ratio': "Ratio de mots vides (stopwords)", 'mean_polarity_diff': "Polarité moyenne (positif/négatif)", 'syntactic_complexity_score': "Complexité syntaxique (subordonnées)", 'ratio_interrogative': "Ratio de phrases interrogatives", 'ratio_exclamative': "Ratio de phrases exclamatives", 'ratio_declarative': "Ratio de phrases déclaratives", 'imparfait_ratio': "Ratio de verbes à l'imparfait", 'futur_ratio': "Ratio de verbes au futur", 'conditional_ratio': "Ratio de verbes au conditionnel", 'passive_voice_ratio': "Ratio de tournures passives" } friendly_names = [COLS_MAP_V2.get(col, col) for col in stylometric_cols] # Slice features to match coefficients length sty_importances = importances[:len(stylometric_cols)] imp_df = pd.DataFrame({"feature": friendly_names, "importance": sty_importances}) imp_df = imp_df.sort_values(by="importance", ascending=False).head(10).sort_values(by="importance", ascending=True) plt.figure(figsize=(10, 6)) colors = ["#3f51b5"] * len(imp_df) bars = plt.barh(imp_df["feature"], imp_df["importance"], color=colors, alpha=0.85) # Label bars for bar in bars: width = bar.get_width() plt.text(width + 0.001, bar.get_y() + bar.get_height()/2, f'{width:.4f}', va='center', ha='left', fontsize=10, fontweight='bold', color='#333333') plt.title("Top 10 des caractéristiques stylométriques les plus discriminantes (Importance XGBoost)", fontsize=14, fontweight="bold", pad=15) plt.xlabel("Importance relative (Gain de pureté)", fontsize=12) plt.ylabel("Caractéristique", fontsize=12) plt.tight_layout() plot5_path = os.path.join(plots_dir, "top_features.png") plt.savefig(plot5_path, dpi=300) plt.savefig(os.path.join(output_dir, "top_features.png"), dpi=300) plt.close() else: model = pkg["model"] model_key = pkg["model_key"] if "logistic_regression" in model_key or "hybrid" in model_key: coefs = model.coef_[0] cols = pkg["stylometric_cols"] if model_key == "logistic_regression_sty" else (pkg["ngram_cols"] if model_key == "logistic_regression_ng" else pkg["hybrid_cols"]) cols = cols[:len(coefs)] word_vectorizer = joblib.load(pkg["vectorizer_words_path"]) char_vectorizer = joblib.load(pkg["vectorizer_chars_path"]) feature_names = [] for f in cols: if f.startswith("ngram_word_"): idx = int(f.split("_")[-1]) feature_names.append(f"Word: '{word_vectorizer.get_feature_names_out()[idx]}'") elif f.startswith("ngram_char_"): idx = int(f.split("_")[-1]) feature_names.append(f"Char: '{char_vectorizer.get_feature_names_out()[idx]}'") else: feature_names.append(f) coef_df = pd.DataFrame({"feature": feature_names, "coef": coefs}) coef_df = coef_df.sort_values(by="coef", ascending=False) top_ai = coef_df.head(5) top_human = coef_df.tail(5) top_plot = pd.concat([top_ai, top_human]).sort_values(by="coef", ascending=True) plt.figure(figsize=(10, 6)) colors = ["#3f51b5" if val < 0 else "#e91e63" for val in top_plot["coef"]] bars = plt.barh(top_plot["feature"], top_plot["coef"], color=colors, alpha=0.85) plt.axvline(x=0, color="#222222", linewidth=1.0) for bar in bars: width = bar.get_width() label_x = width + (0.05 if width >= 0 else -0.55) align = 'left' if width >= 0 else 'right' plt.text(label_x, bar.get_y() + bar.get_height()/2, f'{width:.3f}', va='center', ha=align, fontsize=10, fontweight='bold', color='#333333') plt.title("Top 10 des caractéristiques discriminantes (Coefficients de Régression)", fontsize=14, fontweight="bold", pad=15) plt.xlabel("Importance du coefficient (négatif = Humain, positif = IA)", fontsize=12) plt.ylabel("Caractéristique", fontsize=12) plt.text(0.95, 0.05, "Indique Style IA →", transform=plt.gca().transAxes, color="#e91e63", fontweight="bold", ha="right", fontsize=11) plt.text(0.05, 0.05, "← Indique Style Humain", transform=plt.gca().transAxes, color="#3f51b5", fontweight="bold", ha="left", fontsize=11) plt.tight_layout() plot5_path = os.path.join(plots_dir, "top_features.png") plt.savefig(plot5_path, dpi=300) plt.savefig(os.path.join(output_dir, "top_features.png"), dpi=300) plt.close() # --- Plotly Interactive HTML Map --- print("Generating Plotly Interactive Explorer (interactive_map.html)...") import plotly.express as px # Create copy of dataframe to avoid modifying original df_plotly = df.copy() df_plotly["date_str"] = df_plotly["date"].dt.strftime("%Y-%m-%d") # Epsilon for log scale df_plotly["prob_ai_plot"] = df_plotly["prob_ai"] + 1e-4 # Clean text for hover tooltip (max 200 chars) df_plotly["snippet"] = df_plotly["text"].apply(lambda t: (t[:200] + "...") if isinstance(t, str) and len(t) > 200 else str(t)) # Map predictions to human readable labels df_plotly["type_predit"] = df_plotly["prediction"].map({0: "Humain", 1: "IA"}) fig_interactive = px.scatter( df_plotly, x="date", y="prob_ai_plot", color="party", size="confidence_score", opacity=0.6, hover_data={ "date_str": True, "speaker": True, "party": True, "document_type": True, "prob_ai": ":.4f", "confidence_score": ":.4f", "type_predit": True, "snippet": True, "date": False, "prob_ai_plot": False }, labels={ "date": "Date de l'intervention", "prob_ai_plot": "Score de suspicion (Échelle Log)", "party": "Groupe politique", "confidence_score": "Confiance du modèle", "speaker": "Député / Orateur", "document_type": "Type de document", "prob_ai": "Probabilité d'IA", "type_predit": "Classification", "date_str": "Date", "snippet": "Extrait du texte" }, title="Explorateur Interactif de Détection d'IA dans les Débats Parlementaires Français (2004-2026)" ) # Customize layout and add custom log ticks fig_interactive.update_layout( template="plotly_white", yaxis=dict( type="log", tickvals=[1e-4, 1e-3, 1e-2, 1e-1, 1.0], ticktext=["0.0001 (Humain)", "0.001", "0.01", "0.1", "1.0 (IA)"], title="Score de suspicion d'IA (Échelle Log)" ), xaxis=dict(title="Date de l'intervention"), hoverlabel=dict( bgcolor="white", font_size=12, font_family="Arial" ), legend_title_text="Groupe politique" ) # Save the interactive chart as HTML html_out_path = os.path.join(output_dir, "interactive_map.html") fig_interactive.write_html(html_out_path) # Save to plots dir as well fig_interactive.write_html(os.path.join(plots_dir, "interactive_map.html")) print(f"Plotly interactive map saved to {html_out_path}") print(f"All plots saved in {plots_dir} and duplicate copies in {output_dir}.") if __name__ == "__main__": main()